Deep Learning CT Denoising for Structured and Unstructured Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical images, particularly low-dose CT scans, suffer from both structured and unstructured noise, which degrade image quality and hinder accurate diagnosis, while high-dose scans expose patients to excessive radiation.
Innovation Solution
A multi-stage deep learning method trains a noise reduction neural network to identify and reduce both structured and unstructured noise by alternating input and target datasets with varying noise levels, followed by retraining with ultra-low noise data equivalents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If low-dose CT scan is used to reduce radiation exposure, then patient safety is improved, but image quality deteriorates due to higher noise levels
Solution Approach 1:
The patent converts the harmful effect of noise in low-dose images into a beneficial training signal for the deep learning network. By treating noisy low-dose images as training data to learn denoising patterns, the system transforms the presence of noise from a degradation factor into a learning opportunity, enabling the network to recognize and remove noise patterns while preserving diagnostic features.
Solution Approach 2:
The patent performs preliminary denoising actions during the training phase by preparing training pairs that include noisy low-dose images and corresponding clean reference images. The network is pre-trained to recognize noise patterns and learn the mapping from noisy to clean images before actual clinical use, enabling rapid denoising of new low-dose images without requiring real-time reference scans.
2Manufacturing precision
If high-dose CT scan is used to improve image quality, then diagnostic clarity is improved, but radiation exposure increases
Solution Approach 1:
The patent creates synthetic training pairs by copying and transforming low-dose images to simulate high-dose reference images. Instead of requiring actual high-dose scans for training, the system generates synthetic high-dose equivalents through image processing and augmentation techniques, allowing the network to learn from copied transformations rather than requiring additional radiation exposure during data collection.
Solution Approach 2:
The patent changes the noise level parameter of training images to create a spectrum of training data ranging from very noisy to relatively clean. By systematically varying noise parameters and training the network on this graduated scale, the system learns to progressively reduce noise levels, enabling it to transform extremely noisy low-dose images into high-quality images comparable to high-dose scans.
3Manufacturing precision
If deep learning network is trained with noisy data to reduce unstructured noise, then unstructured noise removal is improved, but structured noise remains challenging
Solution Approach 1:
The patent segments the noise reduction task into distinct phases: first training on unstructured noise patterns using standard denoising approaches, then separately addressing structured noise through additional training with diverse artifacts. This segmentation allows the network to specialize in different noise types sequentially, building comprehensive denoising capability through staged training rather than attempting to handle all noise types simultaneously.
Solution Approach 2:
The patent implements dynamic training where the network's training objectives and data distributions are adaptively adjusted during training. The system transitions from static training datasets to dynamic curricula that introduce increasingly complex noise patterns and artifact types, allowing the network to progressively adapt to different noise characteristics and improve its handling of both unstructured and structured noise.
4Manufacturing precision
If deep learning network is retrained with ultra-low noise data equivalents, then structured noise reduction is improved, but training complexity increases
Solution Approach 1:
The patent performs preliminary generation of ultra-low noise data equivalents before the second training stage. By pre-processing and preparing these enhanced training targets in advance, the system avoids the complexity of generating them in real-time during training. The pre-prepared data equivalents serve as fixed reference targets, simplifying the training process while still enabling the network to learn sophisticated structured noise removal patterns.
Solution Approach 2:
The patent introduces ultra-low noise data equivalents as an intermediary training target that bridges the gap between noisy input images and ideal clean outputs. These intermediaries serve as simplified training targets that capture essential structural information while removing complex noise patterns, making the learning task more tractable and reducing training complexity compared to using actual high-dose clinical images as targets.
Data Source
AI summary
A multi-stage deep learning method is provided, comprising training a deep learning network using a set of training pairs interchangeably including input data from a first noisy dataset with a first noise level and target data from a second noisy dataset with a second noise level, and input data from the second noisy dataset and target data from the first noisy dataset; generating an ultra -low noise data equivalent based on a low noise data fed into the trained deep learning network; and retraining the deep learning network on the set of training pairs using the target data of the set of training pairs in a first retraining step, and using the ultra-low noise data equivalent as target data in a second retraining step.